Tian Yuandong (Mrs. Tian Yuandong)

Tian Yuandong's team teamed up with researchers at the University of California, Berkeley, San Diego, to use the "superposition state" in continuous thinkingTian Yuandong, allowing LLMs to conduct parallel reasoning, and determining the reachability of directed graphsTian YuandongIn order to give the Transformer complex reasoning capabilities, Meta FAIR Tian Yuandong's team proposed that Searchformer is a Transformer model for multi-step planning tasks, such as maze navigation and box pushing. It can calculate optimal plans, and the number of search steps required is much less than that of traditional symbolic planning algorithm such as A #8727 search. The team proposed a dynamic search guidance method to achieve this. First, train a Transform.

Heavy dry goods were delivered to author Tian Yuandong as soon as possibleTian YuandongTian Yuandong led scientists from UC Berkeley and UCSD to use "superimposed states" in continuous space to allow large models to conduct parallel reasoning, greatly improving the model's reachability in graphs.

The birth of the new king of open source models has brought unprecedented opportunities and challenges. Karpathy, the great AI god, predicts that closed-source models will soon catch up. Jim Fan, a senior scientist at Nvidia, said that the power of GPT4 is already in our hands. This is a historic moment. Meta researcher Tian Yuandong announced the beginning of a new era. Open source LLM is on par with closed-source LLM, and even better than Nvidia senior scientist Jim.

Professor Tian Yuandong

1. Generating long and coherent stories is a challenge in the field of artificial intelligence. Previous models were difficult to generate short stories of more than a few paragraphs, which is far from the length of ordinary short stories. Tian Yuandong's team released a story generator framework called Re3 on EMNLP #3922. The Re3 framework uses a hierarchical generation method to build story outlines and character attributes from the planning stage to the drafting stage.

2. Recently, Tian Yuandong was invited by Horizon's former colleagues on Facebook to go to China to give a lecture hall. He shared topics about games and enhanced learning.

3. In contrast, Baidu, as a company with a broad business layout, provides more diverse development opportunities. For technical talents like Lou Tiancheng, Baidu's investment in the fields of algorithms and artificial intelligence, as well as cooperation with top scholars such as Tian Yuandong, may provide him with a better space to play. In addition, Baidu's international layout, especially its business in the United States, also provides Lou Tiancheng with exchanges and cooperation with the world's top technology companies.

4. Tian Yuandong, developer of the American Facebook company's "Dark Forest" Go software, published an analysis article online saying that "'Alpha Go 'system mainly consists of several parts: a Go Network Policy Network. Given the current situation, predict and sample the next step. The second move of the move Fast Rollows out. The goal is the same as the Go Network, but under the conditions of appropriately sacrificing the quality of the move, the speed is 1000 times faster than the Go Network.

5. In short, Tian Yuandong and others successfully lowered the training threshold for large language models on limited memory devices through the innovative method of GaLore, paving the way for the application of AI technology on a wider range of devices.

Tian Yuandong's team Searchformer became popular

1. Source Zhihu Author Tian Yuandong OpenAI released GPT4 yesterday. It was really good to try it out. In terms of story generation, which I mainly care about, it was generated with ChatGPT.

2. The winners of the Outstanding Thesis Nomination Award include Lu Yucheng, Will Grathwohl, Tian Yuandong and Lorenz Richter, who made outstanding contributions in fields such as best complexity discrete distribution, scalable sampling, understanding non-contrast self-supervised learning dynamics in decentralized training, and using tensor chains to solve high-dimensional parabolic PDEs. The number of Chinese authors shortlisted for this year's ICML is about 430, including 209 Chinese students for one work, a significant increase from last year. Wang Zhaoran and Yang Zhuo.

3. Author Tian Yuandong's Facebook Artificial Intelligence Laboratory qubits have been authorized to edit and publish this article. The author Tian Yuandong graduated from the Department of Robotics at Carnegie Mellon University and currently works at.

4. Alpha Go uses many new technologies, such as neural networks deep learning Monte Carlo tree search method, etc., which has made its strength a substantial leap forward. Tian Yuandong, developer of Facebook's "Dark Forest" Go software in the United States, published an analysis on the Internet. The article said that the Alpha Go system is mainly composed of several parts: 1 Policy Network Policy Network, which predicts and samples the next move given the current situation and 2 Fast rollover.

5. AI Daniutian Yuandong publicly stated that Q* can only solve entry-level mathematical problems, and AGI has a high probability that it cannot realize Q* conjecture through synthetic data. It continues to be popular in the AI community, and everyone is guessing.

6. C Xinzhiyuan Recommended Author Tian Yuandong Xinzhiyuan Introduction At ICML2016, 9 papers from Google DeepMind were accepted, and Xinzhiyuan specially invited the Facebook Artificial Intelligence Laboratory.